What to check when an AI service is down

1. What this page is

Every AI company runs its own status page: Anthropic at status.claude.com, OpenAI at status.openai.com, Google on its Cloud and Workspace dashboards. When something breaks, that is where the company writes it first. The trouble is that they are scattered. If you use several tools you end up visiting three or four pages, and in between you wonder whether the problem is on your side.

This app gathers the data those official pages publish, once a minute, onto one screen, and lists the services that are operational as "Working right now". We measure nothing ourselves; we only relay what each company wrote on its own page. The official page is the source of truth and this screen is a window onto it.

2. What the status words mean

Providers use different vocabularies, so we fold them into five words. Here is the mapping.

• Operational: Statuspage indicator none, Instatus status UP, or no open incident on a Google dashboard.
• Partial outage: indicator minor, a component in degraded_performance or partial_outage, Instatus HASISSUES, or a Google incident whose severity is not high.
• Outage: indicator major or critical, a component in major_outage, or a Google incident with severity high.
• Maintenance: announced maintenance (maintenance, under_maintenance, UNDERMAINTENANCE). No alert is sent for it.
• Could not check: we failed to fetch the page (no response, over 8 seconds, or an unknown value). It says nothing about whether the service is up.

GitHub Copilot is read from only the components whose name contains Copilot, not the whole GitHub page. The Gemini app and the Gemini API are two rows because Google reports them on two different dashboards.

3. What we check and what we cannot

These are the sources checked automatically. Every URL was verified to respond on 4 September 2026. The scope column is what the company itself says the page covers.

Some we cannot check: Grok (xAI), Mistral, DeepSeek, Hugging Face, Together AI, Midjourney publish no machine-readable data, so we only link to them. We do not guess their status.

The interval is 60 seconds: our server fetches once a minute and your screen reads that copy, so a change on a provider page can take up to a minute to show here. Alerts are browser notifications and arrive only while this tab is open. There is no server push, so closing the tab stops them. That is deliberate: it runs with no server cost and avoids an alert that looks like it works and then silently does not.

4. The order to check during an outage

1. Look at that service's row. If it says outage or partial outage, it is the provider. The official page link shows which features are affected and whether a fix is under way. This settles most cases.
2. If it is operational but fails only for you, read the error code. 429 means too many requests, usually your account's rate or quota limit (RFC 6585). 503 means the server is temporarily unavailable (RFC 9110). 529 is the code the Anthropic API returns when it is overloaded (overloaded_error). 401 and 403 are key or permission problems, not outages.
3. Check the infrastructure rows. If Cloudflare, Vercel or Supabase show trouble, many services built on them wobble together. When several things break at once, suspect this first.
4. Rule out your side. If other sites are slow too, it is your network. Corporate networks and VPNs sometimes block one service only. Browser extensions, ad blockers in particular, can break a chat page. Trying a private window or another device usually settles it.
5. Wait a few minutes. For a provider outage there is nothing to do but wait until the incident on the official page moves to monitoring or resolved. With alerts on, you are told when it comes back too.

5. How to pick an alternative

• Pick from the same row. "Working right now" lists only confirmed operational services. If you were using a chat, choose from the models and chat row; if an editor, from the coding tools row.
• If your tool lets you choose the model, just switch models. Editors and tools that offer several providers let you keep your history and settings and only change the model behind them.
• For APIs, try retry and backoff first. 429 and 529 often clear after a short wait. Switching providers means re-tuning prompts and output formats, which costs more.
• Operational does not mean identical. It means the company has not written a problem on its page. Regional slowness or per-model limits often never reach a status page.

6. Where people get stuck

• No alert arrived. The tab was closed, the browser blocked notifications for this site, or the OS is in do-not-disturb. Keep the tab open and check the site settings next to the address bar. Mobile browsers usually stop checking when the screen is off.
• "Could not check" keeps showing. That provider's page is not answering our server. Open the official link directly. If many rows show it at once, the problem is more likely on our server.
• The official page differs from this one. We can lag by up to a minute. If they still differ, the official page is right.
• Why two Gemini rows. Google reports the Gemini app on the Workspace dashboard and the Gemini API on the Cloud dashboard. Each row only reacts when that product appears in that dashboard's incident list.
• Grok, Mistral and DeepSeek are missing. They publish no machine-readable status, so they are link only. They will be added if that changes.